A2 · Oversold Sympathy

A bounce setup: a stock dragged down with its sector by more than its own beta explains, with no company-specific reason for the extra drop.

The short code A2 is display metadata only. The canonical identifier this archetype is keyed by everywhere — config, the screener_candidates.archetype column, the scorer emit — is the descriptive string oversold_sympathy.

Thesis

When a sector sells off, the whole group goes down together regardless of individual merit. Index funds, sector ETFs, and risk-parity programs sell the basket, not the names. A stock with a beta of 1.2 should fall 1.2× its sector’s move — but baskets routinely overshoot, dragging down names that had no business falling as far as they did. The edge is the borrowed drop: the portion of a stock’s decline that its sector beta cannot account for, when no idiosyncratic catalyst justifies the extra pain. That unexplained slice is the mean-reversion candidate.

Direction

LONG. It buys the part of the decline that the market over-attributed.

What it looks for

The evaluator (evaluate_oversold_sympathy) screens on a down day plus a quality pass, then runs a real attribution layer; thresholds from screener_config.json § archetypes.a2:

• Down day — ret_1d <= -3% (ret_1d_floor_pct=-3.0).
• Borrowed drop — borrowed_drop_pct = own_ret_pct − beta_60d × sector_ret_pct, fired when its magnitude exceeds borrowed_drop_floor=0.5 percentage points. This is the drop the sector move doesn’t explain.
• Attribution — attribute_catalyst checks the available drivers within an attribution_earnings_window_days=10 window and assigns a cause. Read the confidence inversion carefully: low attribution confidence (we checked everything and found nothing) is the GOOD outcome — a genuinely unexplained idiosyncratic drop is the real high-conviction A2. High attribution confidence means the drop is explained, so the name is demoted.

How it scores

The score is abs(borrowed_drop_pct) — the larger the unexplained slice, the higher the rank. The state machine is driven by attribution, not just the legs:

• none — quality fail, no down day, or the drop is explained by a structural cause (analyst_cluster_down or structural_derate). Those are slow-bleed falling knives and must fall off the board entirely (the EPAM acceptance test).
• partial — the drop is partly explained by a weaker cause (recent_earnings, an A1 concern; or sector_move, non-idiosyncratic), or it’s unexplained but borrowed_drop and the newsless flag don’t both fire.
• primed — genuinely unexplained idiosyncratic drop (identified_catalyst is None, confidence low) where both the borrowed-drop and newsless flags fire. The firing candidate ranked toward S2.

Pitfalls / when it fails

• The news-coverage blind spot. The has_own_news flag reads only two feeds (Iran International + FXStreet). Company press releases, downgrades, and pre-announcements are NOT in the store, so has_own_news=False means “our feeds didn’t flag it,” not “no news exists.” The attribution layer is the structural answer — it checks real drivers rather than trusting the absence flag — but feed coverage is still a limit.
• Catching the basket knife. If the whole sector keeps bleeding, the borrowed-drop snaps back only when the sector stabilizes; the setup says nothing about sector timing.
• Beta drift. beta_60d is a trailing estimate; a regime change can make it stale, distorting the residual.

Related

• The Screener System (vision)
• Screener Pipeline & Lifecycle (operations)
• Mechanical reference cards: Score · State · Triggers
• Sibling archetypes: A5 · Mechanical Reversion · A1 · Post-Earnings Overreaction